Reproduce “Global Regulation and Excitation via Attention Tuning for Stereo Matching” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jiahao Li, Xinhong Chen, Zhengmin Jiang, Qian Zhou, Yung-Hui Li, Jianping Wang

The GREAT (Global Regulation and Excitation via Attention Tuning) framework enhances iterative stereo matching algorithms by systematically incorporating global contextual information into cost volume construction. This approach improves disparity prediction in challenging regions like occlusions and textureless surfaces, achieving state-of-the-art performance on multiple benchmarks and enabling comparable accuracy with significantly fewer iterations.

Key insights

  • Explicitly integrating global context, both spatially and along epipolar lines, is crucial for improving the robustness and accuracy of iterative stereo matching in challenging scenarios.
  • A multi-component attention framework, combining spatial geometry propagation, epipolar ambiguity resolution, and cost volume excitation, effectively addresses different types of matching ambiguities.
  • Enriching the cost volume with global context allows iterative methods to achieve high accuracy with substantially fewer iterations, leading to improved efficiency and practicality.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.